Aims: Cardiac digital twins (CDTs) are personalised computational models of cardiac function that can be used to quantify patient-specific pathophysiological mechanisms. The models usually contain parameters that need to be calibrated, which is challenging for complex computer models. Widely used methods, such as Markov chain Monte Carlo (MCMC), can be prohibitively computationally expensive for use in fast-paced clinical settings. In this study, we illustrate the strengths and weaknesses of four fast calibration methods, providing a toolkit for CDT practitioners.
Methods: We compared four fast calibration methods: linear emulators, approximate Bayesian computation (ABC), ABC-localised linear emulators and neural posterior estimators (NPEs), as well as a baseline approach of MCMC with a Gaussian process emulator. We used the Koriakanitis and Shi 0D lumped parameter model of the circulatory system, and estimated posterior distributions for key parameters, including pulmonary arterial resistance and compliance, using right ventricular pressure traces.
Results: NPEs learned parameter posterior distributions closest to the true values. ABC and ABC-localised linear emulators consistently produced reasonable but more diffuse (uncertain) posteriors. Linear emulators often performed comparably but sometimes failed in edge cases. In terms of cost, linear emulators are the computationally cheapest approach, requiring only a matrix inversion to perform calibration. Both ABC methods require data comparison and sorting of the training set. NPEs require a large upfront cost to train a neural network, but can subsequently sample posterior distributions in a fraction of a second. MCMC consistently produced good posterior predictions, but is by far the slowest method on a case-by-case basis.
Conclusion: The selection of a calibration method should depend on the requirements for speed, accuracy, and the available computational power. NPEs and local-linear emulation approaches both offer the possibility of fast, accurate calibration.